Combined angular margin and cosine margin softmax loss for music classification based on spectrograms

Spectrograms provide rich feature information of music data. Significant progress has been made in music classification using spectrograms and Convolutional Neural Networks (CNNs). However, the softmax loss commonly used in existing CNNs lacks sufficient power to discriminate deep features of music. To overcome this limitation, we propose a Combined Angular Margin and Cosine Margin Softmax Loss (AMCM-Softmax) approach in this paper to enhance intra-class compactness and inter-class discrepancy simultaneously. Specifically, normalization on the weight vectors and feature vectors is adopted to eliminate radial variations. Then, an angular margin parameter and a cosine margin parameter are introduced to maximize the decision margin by enforcing angular and cosine margin constraints. Consequently, the discrimination of features is enhanced by normalization and margin maximization. The decision boundary and the target logit curve of AMCM-Softmax can provide a clear geometric interpretation. Extensive experiments on music datasets show that AMCM-Softmax consistently outperforms the current state-of-the-art approaches in classifying genre and emotion. Our work also shows that a margin loss function can lead to better performance and be used in an advanced CNN model for music classification.